High-performance and energy-saving autonomous navigation system for aerial-ground robots

Abstract

Aerial-ground robots (AGRs) are increasingly recognized for their valuable roles in search, exploration, and rescue tasks. This is attributed to their exceptional mobility and long endurance, which enable them to seamlessly switch between aerial and ground modes, allowing for hybrid locomotion (i.e., flying and driving) in the above challenging tasks. The success of AGRs relies on an autonomous navigation system comprising a perception module, path planner, and controller. Specifically, the perception module uses sensors (e.g., cameras or lidar) to sense the surrounding environment and build a local map. Next, the path planner searches collision-free paths from this map, and the controller executes these paths adaptively. Additionally, the ability of AGRs to switch modes highlights the need for energy-efficient design to improve their operational efficiency and lifespan for demanding tasks. Given their sensitivity to energy consumption, prioritizing ground paths can help conserve energy.

To achieve high-performance and energy-saving navigation, existing sensor-based AGR navigation systems utilize depth cameras to sense their surroundings and create the Euclidean Signed Distance Field (ESDF) map. This map supports quick and effective path planning, allowing AGRs to find collision-free paths. While this approach is effective in open and unobstructed environments, it often struggles in cluttered or unknown areas due to the narrow field of view in sensor-based mapping and the path planner's inherent flaws. In these scenarios, both performance and energy efficiency are compromised.

This thesis presents two pioneering navigation systems, AGRNav and HE-Nav, exemplifying high performance and energy efficiency while tackling unique research challenges. To address the limitations of the sensor-based method in unknown and occluded environments, we introduced AGRNav, a tailored autonomous navigation solution for AGRs. The key innovation is integrating a lightweight, predictive-based network, SCONet, which uses contextual information to enable efficient and real-time obstacle prediction in occluded areas. This approach effectively mitigates the perceptual constraints typically faced by traditional sensor-based methods. AGRNav also incorporates a query-based method for map updates with a hierarchical path planner, facilitating rapid updates to grid maps and enabling more energy-efficient path planning.

Next, recognizing the redundancy in constructing ESDF maps in existing sensor-based methods, we introduced HE-Nav, a novel system tailored for AGRs. Initially, HE-Nav introduced LBSCNet for enhanced perception, which excels in predicting obstacle distribution in occluded areas. Subsequently, we proposed the AG-Planner for ESDF-free path planning, which eliminates the need for ESDF map construction, thereby streamlining the navigation process. LBSCNet's superior obstacle prediction capabilities, coupled with AG-Planner's efficient trajectory generation, significantly advance the efficiency of AGR autonomous navigation in complex environments.

In conclusion, the two innovative systems we introduce tackle critical challenges that have traditionally hindered AGRs from attaining high-performance and energy-efficient autonomous navigation in environments with occlusions. Furthermore, through extensive testing in simulated environments and on our custom-designed AGR, we have thoroughly validated our system's superior performance and energy efficiency, demonstrating its comprehensive advantages in navigating intricate scenarios.